Embedded Feature Selection Method for a Network-Level Behavioural Analysis Detection Model

Mohammad Hafiz Mohd Yusof, Mohd Rosmadi, Abdullah Mohd., Carsten MapleInternational Journal of Advanced Computer Science and Applications 2018 · DOI

Abstract

Feature selection in network-level behavioural analysis studies is used to represent the network datasets of a monitored space. However, recent studies have shown that current behavioural analysis methods at the network-level have several issues. The reduction of millions of instances, disregarded parameters, removed similarities of most of the traffic flows to reduce information noise, insufficient number of optimised features and ignore instances which are not an entity are amongst the other issue that have been identified as the main issues contributing to the inability to predict zero-day attacks. Therefore, this paper aims to select the optimal features that will improve the prediction and behavioural analysis. The training dataset will be trained to use the embedded feature selection method which incorporates both the filter and wrapper method. Correlation coefficient, r and weighted score, wj will be used. The accepted or selected features will be optimised uses Beta distribution functions, β, to find its maximum likelihood, Ɩmax. The final selected features will be trained by the Bayesian Network classifier and tested through several testing datasets. Finally, this method was compared to several other feature selection methods. Final results show the proposed selection method’s performance against other datasets consistently outperform other methods.

Introduction

Traditional network intrusion detection systems (NIDS) primarily rely on signature-based detection, which identifies known patterns of malicious activity. While effective for established threats, these systems struggle against "zero-day" attacks—previously unknown vulnerabilities or exploits. Behavioural analysis offers a promising alternative by monitoring network traffic for anomalies. However, current behavioural models often suffer from over-aggressive data reduction, where millions of instances or similar traffic flows are discarded to minimize noise. This process inadvertently removes subtle clues essential for identifying emerging threats. Furthermore, many models lack historical context ("priori" information) and struggle with the discrete nature of network packet data.

This research addresses these challenges by proposing an embedded feature selection method that optimizes data representation rather than simply reducing volume. By selecting the most relevant features and refining them through statistical optimization, the model aims to achieve high detection sensitivity with computational efficiency.

Methodology

The proposed detection model follows a three-phase workflow designed to enhance feature relevance and classification accuracy.

  1. Pre-processing: Instead of standard normalization, the system calculates a "normal score" ($p_{fi}$) for each feature instance. This score helps distinguish between baseline traffic and potential attack vectors.
  2. Embedded Feature Selection: This phase integrates feature selection directly into the training process. It utilizes a hybrid approach:
  3. Filter Method: Employs Pearson correlation ($r$) to identify linear dependencies in numerical data.
  4. Wrapper Method: Uses a weighted score ($w_j$) for nominal data.
  5. Optimization: Selected features are refined using Beta distribution functions ($\beta$) to determine the maximum likelihood ($\ell_{max}$), ensuring the model focuses on the most representative data points.
  6. Classification: The optimized features are processed by a Bayesian Network classifier. This probabilistic model is adept at handling uncertainty and modeling complex, chronological relationships between network variables.
Proposed Feature Selection and Classification Workflow

Experimental Results

The model was evaluated using the NSL-KDD benchmark dataset and a real-world dataset captured from a Malaysian healthcare provider's internet load balancer between August 2016 and August 2017.

Benchmark Performance

On the NSL-40% and KDDTest+_20Percent datasets, the proposed model achieved a 100% detection rate with 86% overall accuracy. In comparison, standard correlation-based methods (using $r > 0$) only achieved an 11% detection rate and suffered from an 89% false alarm rate, highlighting the superiority of the embedded selection approach.

Real-World Zero-Day Prediction

The most significant finding was the model's ability to predict a real-world malware outbreak. By monitoring subtle shifts in baseline traffic, the model detected early indicators of an attack as early as October 2016 (60% detection). By January 2017, detection reached 86% with 100% accuracy—five months before the attack reached its peak in May and June 2017.

Timeline of Real-World Traffic Analysis and Zero-Day Prediction

Conclusion

The embedded feature selection method demonstrates that retaining and optimizing key features is more effective for zero-day detection than aggressive data reduction. While the model achieved a perfect detection rate on several tests, it maintained a 14% False Alarm Rate (FAR) due to "intersection probability" issues where normal traffic mimics attack patterns. Future work will focus on refining the Bayesian Network's logic to better distinguish these edge cases, further reducing false positives without compromising sensitivity.

Summary updated

Abstract

Feature selection in network-level behavioural analysis studies is used to represent the network datasets of a monitored space. However, recent studies have shown that current behavioural analysis methods at the network-level have several issues. The reduction of millions of instances, disregarded parameters, removed similarities of most of the traffic flows to reduce information noise, insufficient number of optimised features and ignore instances which are not an entity are amongst the other issue that have been identified as the main issues contributing to the inability to predict zero-day attacks. Therefore, this paper aims to select the optimal features that will improve the prediction and behavioural analysis. The training dataset will be trained to use the embedded feature selection method which incorporates both the filter and wrapper method. Correlation coefficient, r and weighted score, wj will be used. The accepted or selected features will be optimised uses Beta distribution functions, β, to find its maximum likelihood, Ɩmax. The final selected features will be trained by the Bayesian Network classifier and tested through several testing datasets. Finally, this method was compared to several other feature selection methods. Final results show the proposed selection method’s performance against other datasets consistently outperform other methods.
BibTeX
@article{yusof2018embedded,
  author = {Mohammad Hafiz Mohd Yusof and Mohd Rosmadi and Abdullah Mohd. and Carsten Maple},
  title = {{Embedded Feature Selection Method for a Network-Level Behavioural Analysis Detection Model}},
  journal = {International Journal of Advanced Computer Science and Applications},
  year = {2018},
  doi = {10.14569/ijacsa.2018.091271},
}
Figure